The idea that complex systems emerge from simpler ones is ancient, but the idea that there are rules that govern how this happens is more recent.  In the sciences that deal with life and intelligence (e.g., biology, cognitive science, computer science) our best theories suggest that these changes are evolutionary.  A set of instructions or codes (DNA-based for biological agents) guides a developmental process that builds an agent in a specific environment. That agent then is more or less successful at surviving and making copies of the code that built it.  Accidental or random changes to the code can occur (mutations) in many different ways, so new varieties of agents are constantly being introduced to a population. Working out how this system manages to create the diversity of life on Earth is a major preoccupation of many biologists.  Cognitive scientists depend on these ideas to understand the basic nature of the minds and behaviors they study. And computer scientists borrow these ideas for use in building artificial intelligence. Understanding the nature of evolutionary processes thus seems to be a problem best tackled by integrating all of these perspectives

Among the most difficult puzzles is that some kinds of agents (biological or artificial) seem to be more evolvable than others.  This controversial idea refers to the ability of some populations of agents to adapt more effectively than others when the world around them changes.  The major goal of our project was to better understand the factors that increase or decrease evolvability. Because these kinds of changes occur over hundreds or thousands of generations, they are difficult to study in existing biological systems.  We therefore used populations of simple robots to test key hypotheses about the factors that affect evolvability. We created these robots in both simulated and also physical form, which allowed us to compare the results from these two methods. Our research led to some findings that we expected but many that were surprising.  

Both theory and also some prior research led us to the prediction that modular brain and body plans are more evolvable than non-modular ones. What we discovered was more nuanced. For example, we found that sparsity of connections in a neural network may be the real driver of evolution.  Because sparse networks are more likely to be modular, it may be that sparsity is the real target of selection with modularity coming along for free. Further research shows that it may be misleading to look only at modularity of the controller or ?brain? of the agent. Whether the body plan is modular and how it evolves with the brain also matters.  Even more surprising is that modularity is more likely to evolve in the agent when the environment itself has a modular organization. Bodies, brains, and environments interact in complex ways to shape the course of evolutionary change. Other research from our labs suggests that how evolvable these systems are may depend on factors other than basic genetics.  Epigenetic factors (which genes get turned off and on for a given agent) and developmental processes (especially whether mistakes are made during development) both make contributions to how evolvable a population will be.  

We also began a deeper examination of the properties of the genome, the system that carries information about how to build an agent.  In biological agents this is a complex DNA-based system with many coding features that remain poorly understood. In particular, we want to understand how gene duplication errors contribute to the emergence of modularity and thus evolvability.  Gene duplication may involve segments of a genome that vary from very small to the whole structure. When duplication happens, mutations to one of the copies leaves the other with its original functions. That allows one copy to mutate and perform as its own distinct module.  This should then make the agent more adaptable in the face of changes to its environment. We developed two artificial genomic coding systems that have the right kinds of developmental properties for exploration of these ideas. Research based on this new artificial G-to-P mapping system is ongoing.  

In addition to these substantive research findings, this grant also supported the completion of two Ph.D.s and contributed significantly to the educational experiences of 21 undergraduates.  Most of these undergraduates have been inspired to pursue further work in this area either in graduate school or in industry. The grant has also been acknowledged in 32 refereed publications, three book chapters, one book and and in 30 public talks and presentations. 

Last Modified: 01/26/2020
Modified by: Kenneth R Livingston